13 rows match MATH-500 / AIME 2026 items when LaTeX formatting is ignored
Hi, thanks for recording the decontamination run in such detail. It made this easy to check.
In revision a404865, 13 rows of data/train-00000-of-00001.parquet contain a MATH-500 problem (8 problems) or an AIME 2026 problem (1 problem) in question. The card reports 0 canonical overlap with cs-giung/math-evals@07268cc. We compared these rows with that revision's own math500 and aime26 splits, whose copies are identical to HuggingFaceH4/MATH-500 and MathArena/aime_2026:
- 10 rows have exactly the same letter/digit word sequence as the eval copy. Only LaTeX delimiters, spacing or markup differ (for example
$$...$$vs\[...\], or[[word]]). - 3 rows contain the complete problem plus an added answer-format sentence; in one of them, one word is also spelled differently.
A comparison on lowercase letter/digit tokens that also checks containment, not only whole-row equality, finds all 13. The rows come through DAPO-Math-17k (6), Open-Reasoner-Zero 72k (3), Open-Reasoner-Zero 57k (2) and DeepScaleR (2), according to source.
13 rows: row offset in the train parquet, benchmark, item ID, match
4766 HuggingFaceH4/MATH-500 test/precalculus/285.json same words; delimiters/markup differ *
6304 MathArena/aime_2026 problem_idx 1 same words; delimiters/markup differ
62752 HuggingFaceH4/MATH-500 test/intermediate_algebra/90.json same words; delimiters/markup differ
66590 HuggingFaceH4/MATH-500 test/geometry/229.json same words; delimiters/markup differ
67039 HuggingFaceH4/MATH-500 test/intermediate_algebra/1365.json same words; delimiters/markup differ
73866 HuggingFaceH4/MATH-500 test/intermediate_algebra/582.json same words; delimiters/markup differ
73929 HuggingFaceH4/MATH-500 test/intermediate_algebra/90.json same words; delimiters/markup differ
81979 HuggingFaceH4/MATH-500 test/algebra/1282.json same words; delimiters/markup differ
83024 HuggingFaceH4/MATH-500 test/geometry/561.json one word spelled differently + added answer-format sentence
83057 HuggingFaceH4/MATH-500 test/algebra/1282.json complete problem + added answer-format sentence
83073 HuggingFaceH4/MATH-500 test/counting_and_probability/430.json complete problem + added answer-format sentence
105168 HuggingFaceH4/MATH-500 test/intermediate_algebra/1365.json same words; delimiters/markup differ
105191 HuggingFaceH4/MATH-500 test/intermediate_algebra/582.json same words; delimiters/markup differ
* = this MATH-500 problem also has a near-identical copy in the MATH train split.
Our benchmark set was MATH-500, AIME 2024-2026, GSM8K test and HumanEval. At this threshold, only MATH-500 and AIME 2026 matched. We did not check the other math-evals splits (AMC, OlympiadBench and others).
How we checked: 13-word n-gram overlap (at least 90% of the benchmark item's 13-grams in one row), then a text comparison of each listed row against the pinned file. This finds verbatim and near-verbatim copies only; paraphrases are missed. We have not trained or evaluated any model on this data and make no claim about any score.
CyberNative AI LLC (research). Corrections or questions: hello@cybernative.ai
Thanks for pointing these out. I’ve applied further decontamination, removing 12 of the 13 reported rows along with additional benchmark overlaps. Row 4766 was retained because it changes (+z^2) to (-z^2), rather than only formatting. The update is available in revision 2eabca2.
Thanks for checking, and you're right about row 4766: it has -z^2 where MATH-500 test/precalculus/285 has +z^2, so it is a different problem, not a formatting copy. Our word-level comparison dropped operators and we mislabeled it. We've updated our labeling to check operator differences before calling a match a formatting difference. Thanks for the quick follow-up.